Accurate daily streamflow estimates are crucial for water resources management. Yet, many regions lack high‐temporal‐resolution data due to limited monitoring infrastructure, often relying on monthly aggregates or intermittent observations. Predicting streamflow in these sparsely sampled watersheds remains challenging. This study proposes a deep learning‐based approach using Long Short‐Term Memory, leveraging its inherent advantages in learning long‐term dependencies within hydrological variables and processes to enhance streamflow predictions in sparsely sampled watersheds. The approach was evaluated for simulating daily flow patterns from monthly aggregated and monthly or weekly intermittent observations in two contrasting hydrological settings: near‐natural and human‐influenced watersheds. Results showed that the proposed approach reliably predicts daily flows from monthly aggregates with a median Nash‐Sutcliffe efficiency (NSE) of 0.61 for near‐natural and 0.48 for human‐influenced watersheds. The proposed approach performed even better for daily flow predictions from monthly or weekly intermittent observation, achieving a median NSE of 0.70 and 0.55 for near‐natural and human‐influenced watersheds, respectively. The proposed approach remained robust across different seasons and hydrological regimes, with a median percentage bias of ±5%, except in arid regions. Moreover, data sensitivity analysis indicated that data from wet seasons were crucial for improving model predictions and that weekly data could yield results comparable to daily observations. Overall, this study demonstrates that the deep learning‐based approach offers a robust and accurate representation of daily streamflow patterns from aggregated or intermittent observations, providing valuable hydrological insights and promising solutions for improving water resource management in regions with limited monitoring infrastructures.
